AI's cheapest model gets four times dearer today, and Google halves its own

Two prices move in opposite directions

From today, the model that made cheap AI famous costs more than four times what it did yesterday. DeepSeek’s V4 Pro rises from $0.87 to $3.96 per million output tokens at peak hours [1]. Tokens are the chunks of text a model reads and writes. A million output tokens is roughly 700,000 words, so this is the price of the machine’s writing, not your reading.

Off-peak, the same model costs $1.98 [1]. The lighter V4 Flash goes from $0.28 to $1.32 at peak, $0.66 off-peak [1]. DeepSeek said the peak and off-peak split lets it “allocate resources more reasonably” [1]. That is rationing by the clock, the way electricity is priced. It is also an admission. The low prices were always a promotion, due to expire on 31 May, made permanent in May, and now reversed anyway [1].

DeepSeek is still cheap by comparison. Moonshot’s Kimi K3 charges $15 per million output tokens and OpenAI’s most advanced model $30 [1]. But everyone who built a product on the assumption that Chinese models stay near-free just had their cost base move under them.

Google moved the other way. Gemini 3.7 Flash arrived three weeks after 3.6 Flash, at $0.75 per million input tokens and $3.75 output [2]. That is an introductory rate. On 1 January 2027 it doubles, to $1.50 and $7.50 [2]. So both companies are doing the same thing on different timetables: an introductory price, then the real one.

What Google is selling is worth naming, because it is the quiet cost nobody prices. The pitch for 3.7 Flash is fewer retries and less manual supervision [2]. On Google’s own benchmarks it scores 43.6% on a production-code test, up from 34.4%, and 65.3% on a long-horizon software task, up from 49.0% [2]. Those are the company’s numbers, not an independent lab’s, and the picture is mixed elsewhere: on one terminal benchmark it trails a rival, 85.8% to 87.4% [2]. But the direction is clear. When an agent, meaning software that carries out a multi-step task on its own, gets something wrong, it tries again. Every retry is another bill. Cheap models made retries feel free. They are not free any more.

Nvidia pushed from the opposite end, releasing Nemotron 3.5 Lightning, a free model small enough to run on a single graphics chip in a laptop [3]. Its logic is plain. “Free AI should be great for chips,” Jensen Huang said [3].

The money arrives anyway

Databricks closed $5 billion at a $190 billion valuation [4]. Its chief executive Ali Ghodsi says the company wanted $1 billion and found $15 billion of interest [4]. It is running at $7 billion of annualised revenue, growing 80%, and is cash-flow positive [4]. It has still raised $20 billion in 20 months, because, in his words, AI is expensive [4].

Cognition, which makes the coding agent Devin, is in talks at $40 billion, three months after raising $1 billion at $26 billion [5]. Lovable raised $400 million at $13.3 billion [6]. CodeRabbit, which reviews code, was valued at $1.5 billion [7]. Thrive Holdings raised $2 billion to buy accounting and IT firms and put AI inside them; OpenAI owns a stake and lends it staff [8]. Anthropic’s finance chief is holding early meetings with investors before a possible listing, with no valuation discussed yet [9]. Nvidia signed up six Wall Street firms to raise more than $500 billion for data centres, with an option to backstop up to $125 billion of it itself [10].

Two people counted where that money comes from. More than 70% of Alphabet’s net quarterly income came from the rising value of its stakes in other companies, mostly SpaceX [11]. About 65% of Amazon’s came from its stake in Anthropic [11]. “It’s circular,” said the researcher who ran the numbers, Matt King [11]. The investor Steve Eisman put a figure on the concentration. OpenAI and Anthropic account for roughly 70% of AI-related revenue at Microsoft, Amazon, Google and Oracle, and a quarter to a third of their cloud revenue [12].

When the thing in the middle goes quiet

On Tuesday morning, Petlibro’s servers stopped answering, and cats and dogs went unfed [13][14]. The company makes feeders that dispense food on a schedule. Its chief executive said the outage was about how the app talks to the devices through its servers [13].

The company’s position is that schedules are stored on the feeder itself and run without any internet command [13]. Owners reported otherwise: some feeders missed every meal, some missed a few, some ran late [13][14]. The “Feed Now” button, which does need the servers, stopped working [13][14]. And the status page carries the sentence that matters most. Some records from the outage “may not be recoverable”, including feeding records [13]. An owner away from home could not press the button, and could not look up whether the cat had eaten.

The business version ran the same week. LexisNexis pulled three services offline after unusual activity on servers run by an outside supplier [15]. One customer said they would seek compensation, noting that firms pay tens of thousands of pounds a year and serve their own clients off it [15].

Insurance has now caught up. AIG launched cover that pays out automatically when cloud downtime crosses a preset trigger, built with a firm that monitors more than 750 data centres and 9,000 software providers [16]. That is how a risk becomes official: someone starts selling protection against it. Global business spending on cloud infrastructure hit $129 billion in the first quarter alone, up 35% on a year earlier [16].

The courts move on the platforms

Google began letting rival app stores be downloaded from the Play Store this month, under a 2023 antitrust loss to Epic [17][18]. Judge James Donato looked at how it did it and gave the company one week to fix what he called anticompetitive friction [17]. Searching Play for “app store” or “aptoide”, the first rival store, did not return it, and the button said “view” rather than “install” [17]. “That is not acceptable,” Donato said [17].

Meta lost its attempt to stop a separate case reaching a jury. An appeals court ruled 3-0 that the law shielding platforms from liability for user content is a defence at trial, not a blanket immunity from being sued [19]. Meta’s own filing put the states’ damages claims above $1.4 trillion, and the trial was due to start on 19 August [19].

The break-ins started with a phone call

Attackers dumped 1.6 million RingCentral customers’ email addresses, names, addresses and phone numbers online after the company refused to pay [20]. The gang says it got in by ringing an employee and talking them out of their password [20].

A quieter one reached further. Poisoned versions of a widely used AI development tool scraped credentials out of 2,500 organisations during a 40-minute window in March, including Microsoft, Amazon, Cisco, Samsung and Salesforce [21]. Researchers found them inside a 195-terabyte file [21]. Microsoft’s monthly patch round carried 421 flaws, one of which North Korean attackers had been exploiting since early June [22].

Elsewhere

Accel closed a $550 million India fund in weeks, 19 months after the last one and with more than half of that still unspent [23]. Linus Torvalds says oversized Linux updates are now normal, driven by fixes from AI review tools he is not thrilled about [24]. And X published the ranking code behind its main feed, along with a tool letting users see whether their own posts have been held back [25].